mm/huge_memory.c

   1 /*
   2  *  Copyright (C) 2009  Red Hat, Inc.
   3  *
   4  *  This work is licensed under the terms of the GNU GPL, version 2. See
   5  *  the COPYING file in the top-level directory.
   6  */
   7
   8 #include <linux/mm.h>
   9 #include <linux/sched.h>
  10 #include <linux/highmem.h>
  11 #include <linux/hugetlb.h>
  12 #include <linux/mmu_notifier.h>
  13 #include <linux/rmap.h>
  14 #include <linux/swap.h>
  15 #include <linux/shrinker.h>
  16 #include <linux/mm_inline.h>
  17 #include <linux/kthread.h>
  18 #include <linux/khugepaged.h>
  19 #include <linux/freezer.h>
  20 #include <linux/mman.h>
  21 #include <linux/pagemap.h>
  22 #include <linux/migrate.h>
  23 #include <linux/hashtable.h>
  24
  25 #include <asm/tlb.h>
  26 #include <asm/pgalloc.h>
  27 #include "internal.h"
  28
  29 /*
  30  * By default transparent hugepage support is enabled for all mappings
  31  * and khugepaged scans all mappings. Defrag is only invoked by
  32  * khugepaged hugepage allocations and by page faults inside
  33  * MADV_HUGEPAGE regions to avoid the risk of slowing down short lived
  34  * allocations.
  35  */
  36 unsigned long transparent_hugepage_flags __read_mostly =
  37 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
  38         (1<<TRANSPARENT_HUGEPAGE_FLAG)|
  39 #endif
  40 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
  41         (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
  42 #endif
  43         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)|
  44         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
  45         (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
  46
  47 /* default scan 8*512 pte (or vmas) every 30 second */
  48 static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8;
  49 static unsigned int khugepaged_pages_collapsed;
  50 static unsigned int khugepaged_full_scans;
  51 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
  52 /* during fragmentation poll the hugepage allocator once every minute */
  53 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
  54 static struct task_struct *khugepaged_thread __read_mostly;
  55 static DEFINE_MUTEX(khugepaged_mutex);
  56 static DEFINE_SPINLOCK(khugepaged_mm_lock);
  57 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
  58 /*
  59  * default collapse hugepages if there is at least one pte mapped like
  60  * it would have happened if the vma was large enough during page
  61  * fault.
  62  */
  63 static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1;
  64
  65 static int khugepaged(void *none);
  66 static int khugepaged_slab_init(void);
  67
  68 #define MM_SLOTS_HASH_BITS 10
  69 static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
  70
  71 static struct kmem_cache *mm_slot_cache __read_mostly;
  72
  73 /**
  74  * struct mm_slot - hash lookup from mm to mm_slot
  75  * @hash: hash collision list
  76  * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
  77  * @mm: the mm that this information is valid for
  78  */
  79 struct mm_slot {
  80         struct hlist_node hash;
  81         struct list_head mm_node;
  82         struct mm_struct *mm;
  83 };
  84
  85 /**
  86  * struct khugepaged_scan - cursor for scanning
  87  * @mm_head: the head of the mm list to scan
  88  * @mm_slot: the current mm_slot we are scanning
  89  * @address: the next address inside that to be scanned
  90  *
  91  * There is only the one khugepaged_scan instance of this cursor structure.
  92  */
  93 struct khugepaged_scan {
  94         struct list_head mm_head;
  95         struct mm_slot *mm_slot;
  96         unsigned long address;
  97 };
  98 static struct khugepaged_scan khugepaged_scan = {
  99         .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
 100 };
 101
 102
 103 static int set_recommended_min_free_kbytes(void)
 104 {
 105         struct zone *zone;
 106         int nr_zones = 0;
 107         unsigned long recommended_min;
 108
 109         if (!khugepaged_enabled())
 110                 return 0;
 111
 112         for_each_populated_zone(zone)
 113                 nr_zones++;
 114
 115         /* Make sure at least 2 hugepages are free for MIGRATE_RESERVE */
 116         recommended_min = pageblock_nr_pages * nr_zones * 2;
 117
 118         /*
 119          * Make sure that on average at least two pageblocks are almost free
 120          * of another type, one for a migratetype to fall back to and a
 121          * second to avoid subsequent fallbacks of other types There are 3
 122          * MIGRATE_TYPES we care about.
 123          */
 124         recommended_min += pageblock_nr_pages * nr_zones *
 125                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
 126
 127         /* don't ever allow to reserve more than 5% of the lowmem */
 128         recommended_min = min(recommended_min,
 129                               (unsigned long) nr_free_buffer_pages() / 20);
 130         recommended_min <<= (PAGE_SHIFT-10);
 131
 132         if (recommended_min > min_free_kbytes)
 133                 min_free_kbytes = recommended_min;
 134         setup_per_zone_wmarks();
 135         return 0;
 136 }
 137 late_initcall(set_recommended_min_free_kbytes);
 138
 139 static int start_khugepaged(void)
 140 {
 141         int err = 0;
 142         if (khugepaged_enabled()) {
 143                 if (!khugepaged_thread)
 144                         khugepaged_thread = kthread_run(khugepaged, NULL,
 145                                                         "khugepaged");
 146                 if (unlikely(IS_ERR(khugepaged_thread))) {
 147                         printk(KERN_ERR
 148                                "khugepaged: kthread_run(khugepaged) failed\n");
 149                         err = PTR_ERR(khugepaged_thread);
 150                         khugepaged_thread = NULL;
 151                 }
 152
 153                 if (!list_empty(&khugepaged_scan.mm_head))
 154                         wake_up_interruptible(&khugepaged_wait);
 155
 156                 set_recommended_min_free_kbytes();
 157         } else if (khugepaged_thread) {
 158                 kthread_stop(khugepaged_thread);
 159                 khugepaged_thread = NULL;
 160         }
 161
 162         return err;
 163 }
 164
 165 static atomic_t huge_zero_refcount;
 166 static struct page *huge_zero_page __read_mostly;
 167
 168 static inline bool is_huge_zero_page(struct page *page)
 169 {
 170         return ACCESS_ONCE(huge_zero_page) == page;
 171 }
 172
 173 static inline bool is_huge_zero_pmd(pmd_t pmd)
 174 {
 175         return is_huge_zero_page(pmd_page(pmd));
 176 }
 177
 178 static struct page *get_huge_zero_page(void)
 179 {
 180         struct page *zero_page;
 181 retry:
 182         if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
 183                 return ACCESS_ONCE(huge_zero_page);
 184
 185         zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
 186                         HPAGE_PMD_ORDER);
 187         if (!zero_page) {
 188                 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
 189                 return NULL;
 190         }
 191         count_vm_event(THP_ZERO_PAGE_ALLOC);
 192         preempt_disable();
 193         if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
 194                 preempt_enable();
 195                 __free_page(zero_page);
 196                 goto retry;
 197         }
 198
 199         /* We take additional reference here. It will be put back by shrinker */
 200         atomic_set(&huge_zero_refcount, 2);
 201         preempt_enable();
 202         return ACCESS_ONCE(huge_zero_page);
 203 }
 204
 205 static void put_huge_zero_page(void)
 206 {
 207         /*
 208          * Counter should never go to zero here. Only shrinker can put
 209          * last reference.
 210          */
 211         BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
 212 }
 213
 214 static int shrink_huge_zero_page(struct shrinker *shrink,
 215                 struct shrink_control *sc)
 216 {
 217         if (!sc->nr_to_scan)
 218                 /* we can free zero page only if last reference remains */
 219                 return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
 220
 221         if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
 222                 struct page *zero_page = xchg(&huge_zero_page, NULL);
 223                 BUG_ON(zero_page == NULL);
 224                 __free_page(zero_page);
 225         }
 226
 227         return 0;
 228 }
 229
 230 static struct shrinker huge_zero_page_shrinker = {
 231         .shrink = shrink_huge_zero_page,
 232         .seeks = DEFAULT_SEEKS,
 233 };
 234
 235 #ifdef CONFIG_SYSFS
 236
 237 static ssize_t double_flag_show(struct kobject *kobj,
 238                                 struct kobj_attribute *attr, char *buf,
 239                                 enum transparent_hugepage_flag enabled,
 240                                 enum transparent_hugepage_flag req_madv)
 241 {
 242         if (test_bit(enabled, &transparent_hugepage_flags)) {
 243                 VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags));
 244                 return sprintf(buf, "[always] madvise never\n");
 245         } else if (test_bit(req_madv, &transparent_hugepage_flags))
 246                 return sprintf(buf, "always [madvise] never\n");
 247         else
 248                 return sprintf(buf, "always madvise [never]\n");
 249 }
 250 static ssize_t double_flag_store(struct kobject *kobj,
 251                                  struct kobj_attribute *attr,
 252                                  const char *buf, size_t count,
 253                                  enum transparent_hugepage_flag enabled,
 254                                  enum transparent_hugepage_flag req_madv)
 255 {
 256         if (!memcmp("always", buf,
 257                     min(sizeof("always")-1, count))) {
 258                 set_bit(enabled, &transparent_hugepage_flags);
 259                 clear_bit(req_madv, &transparent_hugepage_flags);
 260         } else if (!memcmp("madvise", buf,
 261                            min(sizeof("madvise")-1, count))) {
 262                 clear_bit(enabled, &transparent_hugepage_flags);
 263                 set_bit(req_madv, &transparent_hugepage_flags);
 264         } else if (!memcmp("never", buf,
 265                            min(sizeof("never")-1, count))) {
 266                 clear_bit(enabled, &transparent_hugepage_flags);
 267                 clear_bit(req_madv, &transparent_hugepage_flags);
 268         } else
 269                 return -EINVAL;
 270
 271         return count;
 272 }
 273
 274 static ssize_t enabled_show(struct kobject *kobj,
 275                             struct kobj_attribute *attr, char *buf)
 276 {
 277         return double_flag_show(kobj, attr, buf,
 278                                 TRANSPARENT_HUGEPAGE_FLAG,
 279                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 280 }
 281 static ssize_t enabled_store(struct kobject *kobj,
 282                              struct kobj_attribute *attr,
 283                              const char *buf, size_t count)
 284 {
 285         ssize_t ret;
 286
 287         ret = double_flag_store(kobj, attr, buf, count,
 288                                 TRANSPARENT_HUGEPAGE_FLAG,
 289                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 290
 291         if (ret > 0) {
 292                 int err;
 293
 294                 mutex_lock(&khugepaged_mutex);
 295                 err = start_khugepaged();
 296                 mutex_unlock(&khugepaged_mutex);
 297
 298                 if (err)
 299                         ret = err;
 300         }
 301
 302         return ret;
 303 }
 304 static struct kobj_attribute enabled_attr =
 305         __ATTR(enabled, 0644, enabled_show, enabled_store);
 306
 307 static ssize_t single_flag_show(struct kobject *kobj,
 308                                 struct kobj_attribute *attr, char *buf,
 309                                 enum transparent_hugepage_flag flag)
 310 {
 311         return sprintf(buf, "%d\n",
 312                        !!test_bit(flag, &transparent_hugepage_flags));
 313 }
 314
 315 static ssize_t single_flag_store(struct kobject *kobj,
 316                                  struct kobj_attribute *attr,
 317                                  const char *buf, size_t count,
 318                                  enum transparent_hugepage_flag flag)
 319 {
 320         unsigned long value;
 321         int ret;
 322
 323         ret = kstrtoul(buf, 10, &value);
 324         if (ret < 0)
 325                 return ret;
 326         if (value > 1)
 327                 return -EINVAL;
 328
 329         if (value)
 330                 set_bit(flag, &transparent_hugepage_flags);
 331         else
 332                 clear_bit(flag, &transparent_hugepage_flags);
 333
 334         return count;
 335 }
 336
 337 /*
 338  * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
 339  * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
 340  * memory just to allocate one more hugepage.
 341  */
 342 static ssize_t defrag_show(struct kobject *kobj,
 343                            struct kobj_attribute *attr, char *buf)
 344 {
 345         return double_flag_show(kobj, attr, buf,
 346                                 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 347                                 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 348 }
 349 static ssize_t defrag_store(struct kobject *kobj,
 350                             struct kobj_attribute *attr,
 351                             const char *buf, size_t count)
 352 {
 353         return double_flag_store(kobj, attr, buf, count,
 354                                  TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 355                                  TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 356 }
 357 static struct kobj_attribute defrag_attr =
 358         __ATTR(defrag, 0644, defrag_show, defrag_store);
 359
 360 static ssize_t use_zero_page_show(struct kobject *kobj,
 361                 struct kobj_attribute *attr, char *buf)
 362 {
 363         return single_flag_show(kobj, attr, buf,
 364                                 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 365 }
 366 static ssize_t use_zero_page_store(struct kobject *kobj,
 367                 struct kobj_attribute *attr, const char *buf, size_t count)
 368 {
 369         return single_flag_store(kobj, attr, buf, count,
 370                                  TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 371 }
 372 static struct kobj_attribute use_zero_page_attr =
 373         __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
 374 #ifdef CONFIG_DEBUG_VM
 375 static ssize_t debug_cow_show(struct kobject *kobj,
 376                                 struct kobj_attribute *attr, char *buf)
 377 {
 378         return single_flag_show(kobj, attr, buf,
 379                                 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 380 }
 381 static ssize_t debug_cow_store(struct kobject *kobj,
 382                                struct kobj_attribute *attr,
 383                                const char *buf, size_t count)
 384 {
 385         return single_flag_store(kobj, attr, buf, count,
 386                                  TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 387 }
 388 static struct kobj_attribute debug_cow_attr =
 389         __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
 390 #endif /* CONFIG_DEBUG_VM */
 391
 392 static struct attribute *hugepage_attr[] = {
 393         &enabled_attr.attr,
 394         &defrag_attr.attr,
 395         &use_zero_page_attr.attr,
 396 #ifdef CONFIG_DEBUG_VM
 397         &debug_cow_attr.attr,
 398 #endif
 399         NULL,
 400 };
 401
 402 static struct attribute_group hugepage_attr_group = {
 403         .attrs = hugepage_attr,
 404 };
 405
 406 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
 407                                          struct kobj_attribute *attr,
 408                                          char *buf)
 409 {
 410         return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
 411 }
 412
 413 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
 414                                           struct kobj_attribute *attr,
 415                                           const char *buf, size_t count)
 416 {
 417         unsigned long msecs;
 418         int err;
 419
 420         err = strict_strtoul(buf, 10, &msecs);
 421         if (err || msecs > UINT_MAX)
 422                 return -EINVAL;
 423
 424         khugepaged_scan_sleep_millisecs = msecs;
 425         wake_up_interruptible(&khugepaged_wait);
 426
 427         return count;
 428 }
 429 static struct kobj_attribute scan_sleep_millisecs_attr =
 430         __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
 431                scan_sleep_millisecs_store);
 432
 433 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
 434                                           struct kobj_attribute *attr,
 435                                           char *buf)
 436 {
 437         return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
 438 }
 439
 440 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
 441                                            struct kobj_attribute *attr,
 442                                            const char *buf, size_t count)
 443 {
 444         unsigned long msecs;
 445         int err;
 446
 447         err = strict_strtoul(buf, 10, &msecs);
 448         if (err || msecs > UINT_MAX)
 449                 return -EINVAL;
 450
 451         khugepaged_alloc_sleep_millisecs = msecs;
 452         wake_up_interruptible(&khugepaged_wait);
 453
 454         return count;
 455 }
 456 static struct kobj_attribute alloc_sleep_millisecs_attr =
 457         __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
 458                alloc_sleep_millisecs_store);
 459
 460 static ssize_t pages_to_scan_show(struct kobject *kobj,
 461                                   struct kobj_attribute *attr,
 462                                   char *buf)
 463 {
 464         return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
 465 }
 466 static ssize_t pages_to_scan_store(struct kobject *kobj,
 467                                    struct kobj_attribute *attr,
 468                                    const char *buf, size_t count)
 469 {
 470         int err;
 471         unsigned long pages;
 472
 473         err = strict_strtoul(buf, 10, &pages);
 474         if (err || !pages || pages > UINT_MAX)
 475                 return -EINVAL;
 476
 477         khugepaged_pages_to_scan = pages;
 478
 479         return count;
 480 }
 481 static struct kobj_attribute pages_to_scan_attr =
 482         __ATTR(pages_to_scan, 0644, pages_to_scan_show,
 483                pages_to_scan_store);
 484
 485 static ssize_t pages_collapsed_show(struct kobject *kobj,
 486                                     struct kobj_attribute *attr,
 487                                     char *buf)
 488 {
 489         return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
 490 }
 491 static struct kobj_attribute pages_collapsed_attr =
 492         __ATTR_RO(pages_collapsed);
 493
 494 static ssize_t full_scans_show(struct kobject *kobj,
 495                                struct kobj_attribute *attr,
 496                                char *buf)
 497 {
 498         return sprintf(buf, "%u\n", khugepaged_full_scans);
 499 }
 500 static struct kobj_attribute full_scans_attr =
 501         __ATTR_RO(full_scans);
 502
 503 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
 504                                       struct kobj_attribute *attr, char *buf)
 505 {
 506         return single_flag_show(kobj, attr, buf,
 507                                 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 508 }
 509 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
 510                                        struct kobj_attribute *attr,
 511                                        const char *buf, size_t count)
 512 {
 513         return single_flag_store(kobj, attr, buf, count,
 514                                  TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 515 }
 516 static struct kobj_attribute khugepaged_defrag_attr =
 517         __ATTR(defrag, 0644, khugepaged_defrag_show,
 518                khugepaged_defrag_store);
 519
 520 /*
 521  * max_ptes_none controls if khugepaged should collapse hugepages over
 522  * any unmapped ptes in turn potentially increasing the memory
 523  * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
 524  * reduce the available free memory in the system as it
 525  * runs. Increasing max_ptes_none will instead potentially reduce the
 526  * free memory in the system during the khugepaged scan.
 527  */
 528 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
 529                                              struct kobj_attribute *attr,
 530                                              char *buf)
 531 {
 532         return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
 533 }
 534 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
 535                                               struct kobj_attribute *attr,
 536                                               const char *buf, size_t count)
 537 {
 538         int err;
 539         unsigned long max_ptes_none;
 540
 541         err = strict_strtoul(buf, 10, &max_ptes_none);
 542         if (err || max_ptes_none > HPAGE_PMD_NR-1)
 543                 return -EINVAL;
 544
 545         khugepaged_max_ptes_none = max_ptes_none;
 546
 547         return count;
 548 }
 549 static struct kobj_attribute khugepaged_max_ptes_none_attr =
 550         __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
 551                khugepaged_max_ptes_none_store);
 552
 553 static struct attribute *khugepaged_attr[] = {
 554         &khugepaged_defrag_attr.attr,
 555         &khugepaged_max_ptes_none_attr.attr,
 556         &pages_to_scan_attr.attr,
 557         &pages_collapsed_attr.attr,
 558         &full_scans_attr.attr,
 559         &scan_sleep_millisecs_attr.attr,
 560         &alloc_sleep_millisecs_attr.attr,
 561         NULL,
 562 };
 563
 564 static struct attribute_group khugepaged_attr_group = {
 565         .attrs = khugepaged_attr,
 566         .name = "khugepaged",
 567 };
 568
 569 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
 570 {
 571         int err;
 572
 573         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 574         if (unlikely(!*hugepage_kobj)) {
 575                 printk(KERN_ERR "hugepage: failed to create transparent hugepage kobject\n");
 576                 return -ENOMEM;
 577         }
 578
 579         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
 580         if (err) {
 581                 printk(KERN_ERR "hugepage: failed to register transparent hugepage group\n");
 582                 goto delete_obj;
 583         }
 584
 585         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
 586         if (err) {
 587                 printk(KERN_ERR "hugepage: failed to register transparent hugepage group\n");
 588                 goto remove_hp_group;
 589         }
 590
 591         return 0;
 592
 593 remove_hp_group:
 594         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
 595 delete_obj:
 596         kobject_put(*hugepage_kobj);
 597         return err;
 598 }
 599
 600 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 601 {
 602         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
 603         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
 604         kobject_put(hugepage_kobj);
 605 }
 606 #else
 607 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
 608 {
 609         return 0;
 610 }
 611
 612 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 613 {
 614 }
 615 #endif /* CONFIG_SYSFS */
 616
 617 static int __init hugepage_init(void)
 618 {
 619         int err;
 620         struct kobject *hugepage_kobj;
 621
 622         if (!has_transparent_hugepage()) {
 623                 transparent_hugepage_flags = 0;
 624                 return -EINVAL;
 625         }
 626
 627         err = hugepage_init_sysfs(&hugepage_kobj);
 628         if (err)
 629                 return err;
 630
 631         err = khugepaged_slab_init();
 632         if (err)
 633                 goto out;
 634
 635         register_shrinker(&huge_zero_page_shrinker);
 636
 637         /*
 638          * By default disable transparent hugepages on smaller systems,
 639          * where the extra memory used could hurt more than TLB overhead
 640          * is likely to save.  The admin can still enable it through /sys.
 641          */
 642         if (totalram_pages < (512 << (20 - PAGE_SHIFT)))
 643                 transparent_hugepage_flags = 0;
 644
 645         start_khugepaged();
 646
 647         return 0;
 648 out:
 649         hugepage_exit_sysfs(hugepage_kobj);
 650         return err;
 651 }
 652 module_init(hugepage_init)
 653
 654 static int __init setup_transparent_hugepage(char *str)
 655 {
 656         int ret = 0;
 657         if (!str)
 658                 goto out;
 659         if (!strcmp(str, "always")) {
 660                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 661                         &transparent_hugepage_flags);
 662                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 663                           &transparent_hugepage_flags);
 664                 ret = 1;
 665         } else if (!strcmp(str, "madvise")) {
 666                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 667                           &transparent_hugepage_flags);
 668                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 669                         &transparent_hugepage_flags);
 670                 ret = 1;
 671         } else if (!strcmp(str, "never")) {
 672                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 673                           &transparent_hugepage_flags);
 674                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 675                           &transparent_hugepage_flags);
 676                 ret = 1;
 677         }
 678 out:
 679         if (!ret)
 680                 printk(KERN_WARNING
 681                        "transparent_hugepage= cannot parse, ignored\n");
 682         return ret;
 683 }
 684 __setup("transparent_hugepage=", setup_transparent_hugepage);
 685
 686 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 687 {
 688         if (likely(vma->vm_flags & VM_WRITE))
 689                 pmd = pmd_mkwrite(pmd);
 690         return pmd;
 691 }
 692
 693 static inline pmd_t mk_huge_pmd(struct page *page, struct vm_area_struct *vma)
 694 {
 695         pmd_t entry;
 696         entry = mk_pmd(page, vma->vm_page_prot);
 697         entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 698         entry = pmd_mkhuge(entry);
 699         return entry;
 700 }
 701
 702 static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
 703                                         struct vm_area_struct *vma,
 704                                         unsigned long haddr, pmd_t *pmd,
 705                                         struct page *page)
 706 {
 707         pgtable_t pgtable;
 708
 709         VM_BUG_ON(!PageCompound(page));
 710         pgtable = pte_alloc_one(mm, haddr);
 711         if (unlikely(!pgtable))
 712                 return VM_FAULT_OOM;
 713
 714         clear_huge_page(page, haddr, HPAGE_PMD_NR);
 715         /*
 716          * The memory barrier inside __SetPageUptodate makes sure that
 717          * clear_huge_page writes become visible before the set_pmd_at()
 718          * write.
 719          */
 720         __SetPageUptodate(page);
 721
 722         spin_lock(&mm->page_table_lock);
 723         if (unlikely(!pmd_none(*pmd))) {
 724                 spin_unlock(&mm->page_table_lock);
 725                 mem_cgroup_uncharge_page(page);
 726                 put_page(page);
 727                 pte_free(mm, pgtable);
 728         } else {
 729                 pmd_t entry;
 730                 entry = mk_huge_pmd(page, vma);
 731                 page_add_new_anon_rmap(page, vma, haddr);
 732                 set_pmd_at(mm, haddr, pmd, entry);
 733                 pgtable_trans_huge_deposit(mm, pgtable);
 734                 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
 735                 mm->nr_ptes++;
 736                 spin_unlock(&mm->page_table_lock);
 737         }
 738
 739         return 0;
 740 }
 741
 742 static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
 743 {
 744         return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp;
 745 }
 746
 747 static inline struct page *alloc_hugepage_vma(int defrag,
 748                                               struct vm_area_struct *vma,
 749                                               unsigned long haddr, int nd,
 750                                               gfp_t extra_gfp)
 751 {
 752         return alloc_pages_vma(alloc_hugepage_gfpmask(defrag, extra_gfp),
 753                                HPAGE_PMD_ORDER, vma, haddr, nd);
 754 }
 755
 756 #ifndef CONFIG_NUMA
 757 static inline struct page *alloc_hugepage(int defrag)
 758 {
 759         return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
 760                            HPAGE_PMD_ORDER);
 761 }
 762 #endif
 763
 764 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
 765                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
 766                 struct page *zero_page)
 767 {
 768         pmd_t entry;
 769         if (!pmd_none(*pmd))
 770                 return false;
 771         entry = mk_pmd(zero_page, vma->vm_page_prot);
 772         entry = pmd_wrprotect(entry);
 773         entry = pmd_mkhuge(entry);
 774         set_pmd_at(mm, haddr, pmd, entry);
 775         pgtable_trans_huge_deposit(mm, pgtable);
 776         mm->nr_ptes++;
 777         return true;
 778 }
 779
 780 int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
 781                                unsigned long address, pmd_t *pmd,
 782                                unsigned int flags)
 783 {
 784         struct page *page;
 785         unsigned long haddr = address & HPAGE_PMD_MASK;
 786         pte_t *pte;
 787
 788         if (haddr >= vma->vm_start && haddr + HPAGE_PMD_SIZE <= vma->vm_end) {
 789                 if (unlikely(anon_vma_prepare(vma)))
 790                         return VM_FAULT_OOM;
 791                 if (unlikely(khugepaged_enter(vma)))
 792                         return VM_FAULT_OOM;
 793                 if (!(flags & FAULT_FLAG_WRITE) &&
 794                                 transparent_hugepage_use_zero_page()) {
 795                         pgtable_t pgtable;
 796                         struct page *zero_page;
 797                         bool set;
 798                         pgtable = pte_alloc_one(mm, haddr);
 799                         if (unlikely(!pgtable))
 800                                 return VM_FAULT_OOM;
 801                         zero_page = get_huge_zero_page();
 802                         if (unlikely(!zero_page)) {
 803                                 pte_free(mm, pgtable);
 804                                 count_vm_event(THP_FAULT_FALLBACK);
 805                                 goto out;
 806                         }
 807                         spin_lock(&mm->page_table_lock);
 808                         set = set_huge_zero_page(pgtable, mm, vma, haddr, pmd,
 809                                         zero_page);
 810                         spin_unlock(&mm->page_table_lock);
 811                         if (!set) {
 812                                 pte_free(mm, pgtable);
 813                                 put_huge_zero_page();
 814                         }
 815                         return 0;
 816                 }
 817                 page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
 818                                           vma, haddr, numa_node_id(), 0);
 819                 if (unlikely(!page)) {
 820                         count_vm_event(THP_FAULT_FALLBACK);
 821                         goto out;
 822                 }
 823                 count_vm_event(THP_FAULT_ALLOC);
 824                 if (unlikely(mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))) {
 825                         put_page(page);
 826                         goto out;
 827                 }
 828                 if (unlikely(__do_huge_pmd_anonymous_page(mm, vma, haddr, pmd,
 829                                                           page))) {
 830                         mem_cgroup_uncharge_page(page);
 831                         put_page(page);
 832                         goto out;
 833                 }
 834
 835                 return 0;
 836         }
 837 out:
 838         /*
 839          * Use __pte_alloc instead of pte_alloc_map, because we can't
 840          * run pte_offset_map on the pmd, if an huge pmd could
 841          * materialize from under us from a different thread.
 842          */
 843         if (unlikely(pmd_none(*pmd)) &&
 844             unlikely(__pte_alloc(mm, vma, pmd, address)))
 845                 return VM_FAULT_OOM;
 846         /* if an huge pmd materialized from under us just retry later */
 847         if (unlikely(pmd_trans_huge(*pmd)))
 848                 return 0;
 849         /*
 850          * A regular pmd is established and it can't morph into a huge pmd
 851          * from under us anymore at this point because we hold the mmap_sem
 852          * read mode and khugepaged takes it in write mode. So now it's
 853          * safe to run pte_offset_map().
 854          */
 855         pte = pte_offset_map(pmd, address);
 856         return handle_pte_fault(mm, vma, address, pte, pmd, flags);
 857 }
 858
 859 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 860                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
 861                   struct vm_area_struct *vma)
 862 {
 863         struct page *src_page;
 864         pmd_t pmd;
 865         pgtable_t pgtable;
 866         int ret;
 867
 868         ret = -ENOMEM;
 869         pgtable = pte_alloc_one(dst_mm, addr);
 870         if (unlikely(!pgtable))
 871                 goto out;
 872
 873         spin_lock(&dst_mm->page_table_lock);
 874         spin_lock_nested(&src_mm->page_table_lock, SINGLE_DEPTH_NESTING);
 875
 876         ret = -EAGAIN;
 877         pmd = *src_pmd;
 878         if (unlikely(!pmd_trans_huge(pmd))) {
 879                 pte_free(dst_mm, pgtable);
 880                 goto out_unlock;
 881         }
 882         /*
 883          * mm->page_table_lock is enough to be sure that huge zero pmd is not
 884          * under splitting since we don't split the page itself, only pmd to
 885          * a page table.
 886          */
 887         if (is_huge_zero_pmd(pmd)) {
 888                 struct page *zero_page;
 889                 bool set;
 890                 /*
 891                  * get_huge_zero_page() will never allocate a new page here,
 892                  * since we already have a zero page to copy. It just takes a
 893                  * reference.
 894                  */
 895                 zero_page = get_huge_zero_page();
 896                 set = set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
 897                                 zero_page);
 898                 BUG_ON(!set); /* unexpected !pmd_none(dst_pmd) */
 899                 ret = 0;
 900                 goto out_unlock;
 901         }
 902         if (unlikely(pmd_trans_splitting(pmd))) {
 903                 /* split huge page running from under us */
 904                 spin_unlock(&src_mm->page_table_lock);
 905                 spin_unlock(&dst_mm->page_table_lock);
 906                 pte_free(dst_mm, pgtable);
 907
 908                 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
 909                 goto out;
 910         }
 911         src_page = pmd_page(pmd);
 912         VM_BUG_ON(!PageHead(src_page));
 913         get_page(src_page);
 914         page_dup_rmap(src_page);
 915         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 916
 917         pmdp_set_wrprotect(src_mm, addr, src_pmd);
 918         pmd = pmd_mkold(pmd_wrprotect(pmd));
 919         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
 920         pgtable_trans_huge_deposit(dst_mm, pgtable);
 921         dst_mm->nr_ptes++;
 922
 923         ret = 0;
 924 out_unlock:
 925         spin_unlock(&src_mm->page_table_lock);
 926         spin_unlock(&dst_mm->page_table_lock);
 927 out:
 928         return ret;
 929 }
 930
 931 void huge_pmd_set_accessed(struct mm_struct *mm,
 932                            struct vm_area_struct *vma,
 933                            unsigned long address,
 934                            pmd_t *pmd, pmd_t orig_pmd,
 935                            int dirty)
 936 {
 937         pmd_t entry;
 938         unsigned long haddr;
 939
 940         spin_lock(&mm->page_table_lock);
 941         if (unlikely(!pmd_same(*pmd, orig_pmd)))
 942                 goto unlock;
 943
 944         entry = pmd_mkyoung(orig_pmd);
 945         haddr = address & HPAGE_PMD_MASK;
 946         if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty))
 947                 update_mmu_cache_pmd(vma, address, pmd);
 948
 949 unlock:
 950         spin_unlock(&mm->page_table_lock);
 951 }
 952
 953 static int do_huge_pmd_wp_zero_page_fallback(struct mm_struct *mm,
 954                 struct vm_area_struct *vma, unsigned long address,
 955                 pmd_t *pmd, pmd_t orig_pmd, unsigned long haddr)
 956 {
 957         pgtable_t pgtable;
 958         pmd_t _pmd;
 959         struct page *page;
 960         int i, ret = 0;
 961         unsigned long mmun_start;       /* For mmu_notifiers */
 962         unsigned long mmun_end;         /* For mmu_notifiers */
 963
 964         page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
 965         if (!page) {
 966                 ret |= VM_FAULT_OOM;
 967                 goto out;
 968         }
 969
 970         if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL)) {
 971                 put_page(page);
 972                 ret |= VM_FAULT_OOM;
 973                 goto out;
 974         }
 975
 976         clear_user_highpage(page, address);
 977         __SetPageUptodate(page);
 978
 979         mmun_start = haddr;
 980         mmun_end   = haddr + HPAGE_PMD_SIZE;
 981         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
 982
 983         spin_lock(&mm->page_table_lock);
 984         if (unlikely(!pmd_same(*pmd, orig_pmd)))
 985                 goto out_free_page;
 986
 987         pmdp_clear_flush(vma, haddr, pmd);
 988         /* leave pmd empty until pte is filled */
 989
 990         pgtable = pgtable_trans_huge_withdraw(mm);
 991         pmd_populate(mm, &_pmd, pgtable);
 992
 993         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
 994                 pte_t *pte, entry;
 995                 if (haddr == (address & PAGE_MASK)) {
 996                         entry = mk_pte(page, vma->vm_page_prot);
 997                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
 998                         page_add_new_anon_rmap(page, vma, haddr);
 999                 } else {
1000                         entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
1001                         entry = pte_mkspecial(entry);
1002                 }
1003                 pte = pte_offset_map(&_pmd, haddr);
1004                 VM_BUG_ON(!pte_none(*pte));
1005                 set_pte_at(mm, haddr, pte, entry);
1006                 pte_unmap(pte);
1007         }
1008         smp_wmb(); /* make pte visible before pmd */
1009         pmd_populate(mm, pmd, pgtable);
1010         spin_unlock(&mm->page_table_lock);
1011         put_huge_zero_page();
1012         inc_mm_counter(mm, MM_ANONPAGES);
1013
1014         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1015
1016         ret |= VM_FAULT_WRITE;
1017 out:
1018         return ret;
1019 out_free_page:
1020         spin_unlock(&mm->page_table_lock);
1021         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1022         mem_cgroup_uncharge_page(page);
1023         put_page(page);
1024         goto out;
1025 }
1026
1027 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
1028                                         struct vm_area_struct *vma,
1029                                         unsigned long address,
1030                                         pmd_t *pmd, pmd_t orig_pmd,
1031                                         struct page *page,
1032                                         unsigned long haddr)
1033 {
1034         pgtable_t pgtable;
1035         pmd_t _pmd;
1036         int ret = 0, i;
1037         struct page **pages;
1038         unsigned long mmun_start;       /* For mmu_notifiers */
1039         unsigned long mmun_end;         /* For mmu_notifiers */
1040
1041         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
1042                         GFP_KERNEL);
1043         if (unlikely(!pages)) {
1044                 ret |= VM_FAULT_OOM;
1045                 goto out;
1046         }
1047
1048         for (i = 0; i < HPAGE_PMD_NR; i++) {
1049                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
1050                                                __GFP_OTHER_NODE,
1051                                                vma, address, page_to_nid(page));
1052                 if (unlikely(!pages[i] ||
1053                              mem_cgroup_newpage_charge(pages[i], mm,
1054                                                        GFP_KERNEL))) {
1055                         if (pages[i])
1056                                 put_page(pages[i]);
1057                         mem_cgroup_uncharge_start();
1058                         while (--i >= 0) {
1059                                 mem_cgroup_uncharge_page(pages[i]);
1060                                 put_page(pages[i]);
1061                         }
1062                         mem_cgroup_uncharge_end();
1063                         kfree(pages);
1064                         ret |= VM_FAULT_OOM;
1065                         goto out;
1066                 }
1067         }
1068
1069         for (i = 0; i < HPAGE_PMD_NR; i++) {
1070                 copy_user_highpage(pages[i], page + i,
1071                                    haddr + PAGE_SIZE * i, vma);
1072                 __SetPageUptodate(pages[i]);
1073                 cond_resched();
1074         }
1075
1076         mmun_start = haddr;
1077         mmun_end   = haddr + HPAGE_PMD_SIZE;
1078         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1079
1080         spin_lock(&mm->page_table_lock);
1081         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1082                 goto out_free_pages;
1083         VM_BUG_ON(!PageHead(page));
1084
1085         pmdp_clear_flush(vma, haddr, pmd);
1086         /* leave pmd empty until pte is filled */
1087
1088         pgtable = pgtable_trans_huge_withdraw(mm);
1089         pmd_populate(mm, &_pmd, pgtable);
1090
1091         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1092                 pte_t *pte, entry;
1093                 entry = mk_pte(pages[i], vma->vm_page_prot);
1094                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1095                 page_add_new_anon_rmap(pages[i], vma, haddr);
1096                 pte = pte_offset_map(&_pmd, haddr);
1097                 VM_BUG_ON(!pte_none(*pte));
1098                 set_pte_at(mm, haddr, pte, entry);
1099                 pte_unmap(pte);
1100         }
1101         kfree(pages);
1102
1103         smp_wmb(); /* make pte visible before pmd */
1104         pmd_populate(mm, pmd, pgtable);
1105         page_remove_rmap(page);
1106         spin_unlock(&mm->page_table_lock);
1107
1108         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1109
1110         ret |= VM_FAULT_WRITE;
1111         put_page(page);
1112
1113 out:
1114         return ret;
1115
1116 out_free_pages:
1117         spin_unlock(&mm->page_table_lock);
1118         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1119         mem_cgroup_uncharge_start();
1120         for (i = 0; i < HPAGE_PMD_NR; i++) {
1121                 mem_cgroup_uncharge_page(pages[i]);
1122                 put_page(pages[i]);
1123         }
1124         mem_cgroup_uncharge_end();
1125         kfree(pages);
1126         goto out;
1127 }
1128
1129 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1130                         unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
1131 {
1132         int ret = 0;
1133         struct page *page = NULL, *new_page;
1134         unsigned long haddr;
1135         unsigned long mmun_start;       /* For mmu_notifiers */
1136         unsigned long mmun_end;         /* For mmu_notifiers */
1137
1138         VM_BUG_ON(!vma->anon_vma);
1139         haddr = address & HPAGE_PMD_MASK;
1140         if (is_huge_zero_pmd(orig_pmd))
1141                 goto alloc;
1142         spin_lock(&mm->page_table_lock);
1143         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1144                 goto out_unlock;
1145
1146         page = pmd_page(orig_pmd);
1147         VM_BUG_ON(!PageCompound(page) || !PageHead(page));
1148         if (page_mapcount(page) == 1) {
1149                 pmd_t entry;
1150                 entry = pmd_mkyoung(orig_pmd);
1151                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1152                 if (pmdp_set_access_flags(vma, haddr, pmd, entry,  1))
1153                         update_mmu_cache_pmd(vma, address, pmd);
1154                 ret |= VM_FAULT_WRITE;
1155                 goto out_unlock;
1156         }
1157         get_page(page);
1158         spin_unlock(&mm->page_table_lock);
1159 alloc:
1160         if (transparent_hugepage_enabled(vma) &&
1161             !transparent_hugepage_debug_cow())
1162                 new_page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
1163                                               vma, haddr, numa_node_id(), 0);
1164         else
1165                 new_page = NULL;
1166
1167         if (unlikely(!new_page)) {
1168                 count_vm_event(THP_FAULT_FALLBACK);
1169                 if (!page) {
1170                         ret = do_huge_pmd_wp_zero_page_fallback(mm, vma,
1171                                         address, pmd, orig_pmd, haddr);
1172                 } else {
1173                         ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
1174                                         pmd, orig_pmd, page, haddr);
1175                         if (ret & VM_FAULT_OOM)
1176                                 split_huge_page(page);
1177                         put_page(page);
1178                 }
1179                 goto out;
1180         }
1181         count_vm_event(THP_FAULT_ALLOC);
1182
1183         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
1184                 put_page(new_page);
1185                 if (page) {
1186                         split_huge_page(page);
1187                         put_page(page);
1188                 }
1189                 ret |= VM_FAULT_OOM;
1190                 goto out;
1191         }
1192
1193         if (!page)
1194                 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1195         else
1196                 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1197         __SetPageUptodate(new_page);
1198
1199         mmun_start = haddr;
1200         mmun_end   = haddr + HPAGE_PMD_SIZE;
1201         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1202
1203         spin_lock(&mm->page_table_lock);
1204         if (page)
1205                 put_page(page);
1206         if (unlikely(!pmd_same(*pmd, orig_pmd))) {
1207                 spin_unlock(&mm->page_table_lock);
1208                 mem_cgroup_uncharge_page(new_page);
1209                 put_page(new_page);
1210                 goto out_mn;
1211         } else {
1212                 pmd_t entry;
1213                 entry = mk_huge_pmd(new_page, vma);
1214                 pmdp_clear_flush(vma, haddr, pmd);
1215                 page_add_new_anon_rmap(new_page, vma, haddr);
1216                 set_pmd_at(mm, haddr, pmd, entry);
1217                 update_mmu_cache_pmd(vma, address, pmd);
1218                 if (!page) {
1219                         add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
1220                         put_huge_zero_page();
1221                 } else {
1222                         VM_BUG_ON(!PageHead(page));
1223                         page_remove_rmap(page);
1224                         put_page(page);
1225                 }
1226                 ret |= VM_FAULT_WRITE;
1227         }
1228         spin_unlock(&mm->page_table_lock);
1229 out_mn:
1230         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1231 out:
1232         return ret;
1233 out_unlock:
1234         spin_unlock(&mm->page_table_lock);
1235         return ret;
1236 }
1237
1238 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1239                                    unsigned long addr,
1240                                    pmd_t *pmd,
1241                                    unsigned int flags)
1242 {
1243         struct mm_struct *mm = vma->vm_mm;
1244         struct page *page = NULL;
1245
1246         assert_spin_locked(&mm->page_table_lock);
1247
1248         if (flags & FOLL_WRITE && !pmd_write(*pmd))
1249                 goto out;
1250
1251         /* Avoid dumping huge zero page */
1252         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1253                 return ERR_PTR(-EFAULT);
1254
1255         page = pmd_page(*pmd);
1256         VM_BUG_ON(!PageHead(page));
1257         if (flags & FOLL_TOUCH) {
1258                 pmd_t _pmd;
1259                 /*
1260                  * We should set the dirty bit only for FOLL_WRITE but
1261                  * for now the dirty bit in the pmd is meaningless.
1262                  * And if the dirty bit will become meaningful and
1263                  * we'll only set it with FOLL_WRITE, an atomic
1264                  * set_bit will be required on the pmd to set the
1265                  * young bit, instead of the current set_pmd_at.
1266                  */
1267                 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
1268                 set_pmd_at(mm, addr & HPAGE_PMD_MASK, pmd, _pmd);
1269         }
1270         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1271                 if (page->mapping && trylock_page(page)) {
1272                         lru_add_drain();
1273                         if (page->mapping)
1274                                 mlock_vma_page(page);
1275                         unlock_page(page);
1276                 }
1277         }
1278         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1279         VM_BUG_ON(!PageCompound(page));
1280         if (flags & FOLL_GET)
1281                 get_page_foll(page);
1282
1283 out:
1284         return page;
1285 }
1286
1287 /* NUMA hinting page fault entry point for trans huge pmds */
1288 int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
1289                                 unsigned long addr, pmd_t pmd, pmd_t *pmdp)
1290 {
1291         struct anon_vma *anon_vma = NULL;
1292         struct page *page;
1293         unsigned long haddr = addr & HPAGE_PMD_MASK;
1294         int page_nid = -1, this_nid = numa_node_id();
1295         int target_nid;
1296         bool page_locked;
1297         bool migrated = false;
1298
1299         spin_lock(&mm->page_table_lock);
1300         if (unlikely(!pmd_same(pmd, *pmdp)))
1301                 goto out_unlock;
1302
1303         page = pmd_page(pmd);
1304         page_nid = page_to_nid(page);
1305         count_vm_numa_event(NUMA_HINT_FAULTS);
1306         if (page_nid == this_nid)
1307                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1308
1309         /*
1310          * Acquire the page lock to serialise THP migrations but avoid dropping
1311          * page_table_lock if at all possible
1312          */
1313         page_locked = trylock_page(page);
1314         target_nid = mpol_misplaced(page, vma, haddr);
1315         if (target_nid == -1) {
1316                 /* If the page was locked, there are no parallel migrations */
1317                 if (page_locked)
1318                         goto clear_pmdnuma;
1319
1320                 /*
1321                  * Otherwise wait for potential migrations and retry. We do
1322                  * relock and check_same as the page may no longer be mapped.
1323                  * As the fault is being retried, do not account for it.
1324                  */
1325                 spin_unlock(&mm->page_table_lock);
1326                 wait_on_page_locked(page);
1327                 page_nid = -1;
1328                 goto out;
1329         }
1330
1331         /* Page is misplaced, serialise migrations and parallel THP splits */
1332         get_page(page);
1333         spin_unlock(&mm->page_table_lock);
1334         if (!page_locked)
1335                 lock_page(page);
1336         anon_vma = page_lock_anon_vma_read(page);
1337
1338         /* Confirm the PTE did not while locked */
1339         spin_lock(&mm->page_table_lock);
1340         if (unlikely(!pmd_same(pmd, *pmdp))) {
1341                 unlock_page(page);
1342                 put_page(page);
1343                 page_nid = -1;
1344                 goto out_unlock;
1345         }
1346
1347         /* Bail if we fail to protect against THP splits for any reason */
1348         if (unlikely(!anon_vma)) {
1349                 put_page(page);
1350                 page_nid = -1;
1351                 goto clear_pmdnuma;
1352         }
1353
1354         /*
1355          * The page_table_lock above provides a memory barrier
1356          * with change_protection_range.
1357          */
1358         if (mm_tlb_flush_pending(mm))
1359                 flush_tlb_range(vma, haddr, haddr + HPAGE_PMD_SIZE);
1360
1361         /*
1362          * Migrate the THP to the requested node, returns with page unlocked
1363          * and pmd_numa cleared.
1364          */
1365         spin_unlock(&mm->page_table_lock);
1366         migrated = migrate_misplaced_transhuge_page(mm, vma,
1367                                 pmdp, pmd, addr, page, target_nid);
1368         if (migrated)
1369                 page_nid = target_nid;
1370
1371         goto out;
1372 clear_pmdnuma:
1373         BUG_ON(!PageLocked(page));
1374         pmd = pmd_mknonnuma(pmd);
1375         set_pmd_at(mm, haddr, pmdp, pmd);
1376         VM_BUG_ON(pmd_numa(*pmdp));
1377         update_mmu_cache_pmd(vma, addr, pmdp);
1378         unlock_page(page);
1379 out_unlock:
1380         spin_unlock(&mm->page_table_lock);
1381
1382 out:
1383         if (anon_vma)
1384                 page_unlock_anon_vma_read(anon_vma);
1385
1386         if (page_nid != -1)
1387                 task_numa_fault(page_nid, HPAGE_PMD_NR, migrated);
1388
1389         return 0;
1390 }
1391
1392 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1393                  pmd_t *pmd, unsigned long addr)
1394 {
1395         int ret = 0;
1396
1397         if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1398                 struct page *page;
1399                 pgtable_t pgtable;
1400                 pmd_t orig_pmd;
1401                 pgtable = pgtable_trans_huge_withdraw(tlb->mm);
1402                 orig_pmd = pmdp_get_and_clear(tlb->mm, addr, pmd);
1403                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1404                 if (is_huge_zero_pmd(orig_pmd)) {
1405                         tlb->mm->nr_ptes--;
1406                         spin_unlock(&tlb->mm->page_table_lock);
1407                         put_huge_zero_page();
1408                 } else {
1409                         page = pmd_page(orig_pmd);
1410                         page_remove_rmap(page);
1411                         VM_BUG_ON(page_mapcount(page) < 0);
1412                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1413                         VM_BUG_ON(!PageHead(page));
1414                         tlb->mm->nr_ptes--;
1415                         spin_unlock(&tlb->mm->page_table_lock);
1416                         tlb_remove_page(tlb, page);
1417                 }
1418                 pte_free(tlb->mm, pgtable);
1419                 ret = 1;
1420         }
1421         return ret;
1422 }
1423
1424 int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1425                 unsigned long addr, unsigned long end,
1426                 unsigned char *vec)
1427 {
1428         int ret = 0;
1429
1430         if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1431                 /*
1432                  * All logical pages in the range are present
1433                  * if backed by a huge page.
1434                  */
1435                 spin_unlock(&vma->vm_mm->page_table_lock);
1436                 memset(vec, 1, (end - addr) >> PAGE_SHIFT);
1437                 ret = 1;
1438         }
1439
1440         return ret;
1441 }
1442
1443 int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1444                   unsigned long old_addr,
1445                   unsigned long new_addr, unsigned long old_end,
1446                   pmd_t *old_pmd, pmd_t *new_pmd)
1447 {
1448         int ret = 0;
1449         pmd_t pmd;
1450
1451         struct mm_struct *mm = vma->vm_mm;
1452
1453         if ((old_addr & ~HPAGE_PMD_MASK) ||
1454             (new_addr & ~HPAGE_PMD_MASK) ||
1455             old_end - old_addr < HPAGE_PMD_SIZE ||
1456             (new_vma->vm_flags & VM_NOHUGEPAGE))
1457                 goto out;
1458
1459         /*
1460          * The destination pmd shouldn't be established, free_pgtables()
1461          * should have release it.
1462          */
1463         if (WARN_ON(!pmd_none(*new_pmd))) {
1464                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1465                 goto out;
1466         }
1467
1468         ret = __pmd_trans_huge_lock(old_pmd, vma);
1469         if (ret == 1) {
1470                 pmd = pmdp_get_and_clear(mm, old_addr, old_pmd);
1471                 VM_BUG_ON(!pmd_none(*new_pmd));
1472                 set_pmd_at(mm, new_addr, new_pmd, pmd);
1473                 spin_unlock(&mm->page_table_lock);
1474         }
1475 out:
1476         return ret;
1477 }
1478
1479 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1480                 unsigned long addr, pgprot_t newprot, int prot_numa)
1481 {
1482         struct mm_struct *mm = vma->vm_mm;
1483         int ret = 0;
1484
1485         if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1486                 pmd_t entry;
1487                 entry = pmdp_get_and_clear(mm, addr, pmd);
1488                 if (!prot_numa) {
1489                         entry = pmd_modify(entry, newprot);
1490                         BUG_ON(pmd_write(entry));
1491                 } else {
1492                         struct page *page = pmd_page(*pmd);
1493
1494                         /* only check non-shared pages */
1495                         if (page_mapcount(page) == 1 &&
1496                             !pmd_numa(*pmd)) {
1497                                 entry = pmd_mknuma(entry);
1498                         }
1499                 }
1500                 set_pmd_at(mm, addr, pmd, entry);
1501                 spin_unlock(&vma->vm_mm->page_table_lock);
1502                 ret = 1;
1503         }
1504
1505         return ret;
1506 }
1507
1508 /*
1509  * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1510  * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1511  *
1512  * Note that if it returns 1, this routine returns without unlocking page
1513  * table locks. So callers must unlock them.
1514  */
1515 int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1516 {
1517         spin_lock(&vma->vm_mm->page_table_lock);
1518         if (likely(pmd_trans_huge(*pmd))) {
1519                 if (unlikely(pmd_trans_splitting(*pmd))) {
1520                         spin_unlock(&vma->vm_mm->page_table_lock);
1521                         wait_split_huge_page(vma->anon_vma, pmd);
1522                         return -1;
1523                 } else {
1524                         /* Thp mapped by 'pmd' is stable, so we can
1525                          * handle it as it is. */
1526                         return 1;
1527                 }
1528         }
1529         spin_unlock(&vma->vm_mm->page_table_lock);
1530         return 0;
1531 }
1532
1533 pmd_t *page_check_address_pmd(struct page *page,
1534                               struct mm_struct *mm,
1535                               unsigned long address,
1536                               enum page_check_address_pmd_flag flag)
1537 {
1538         pmd_t *pmd, *ret = NULL;
1539
1540         if (address & ~HPAGE_PMD_MASK)
1541                 goto out;
1542
1543         pmd = mm_find_pmd(mm, address);
1544         if (!pmd)
1545                 goto out;
1546         if (pmd_none(*pmd))
1547                 goto out;
1548         if (pmd_page(*pmd) != page)
1549                 goto out;
1550         /*
1551          * split_vma() may create temporary aliased mappings. There is
1552          * no risk as long as all huge pmd are found and have their
1553          * splitting bit set before __split_huge_page_refcount
1554          * runs. Finding the same huge pmd more than once during the
1555          * same rmap walk is not a problem.
1556          */
1557         if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1558             pmd_trans_splitting(*pmd))
1559                 goto out;
1560         if (pmd_trans_huge(*pmd)) {
1561                 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1562                           !pmd_trans_splitting(*pmd));
1563                 ret = pmd;
1564         }
1565 out:
1566         return ret;
1567 }
1568
1569 static int __split_huge_page_splitting(struct page *page,
1570                                        struct vm_area_struct *vma,
1571                                        unsigned long address)
1572 {
1573         struct mm_struct *mm = vma->vm_mm;
1574         pmd_t *pmd;
1575         int ret = 0;
1576         /* For mmu_notifiers */
1577         const unsigned long mmun_start = address;
1578         const unsigned long mmun_end   = address + HPAGE_PMD_SIZE;
1579
1580         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1581         spin_lock(&mm->page_table_lock);
1582         pmd = page_check_address_pmd(page, mm, address,
1583                                      PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG);
1584         if (pmd) {
1585                 /*
1586                  * We can't temporarily set the pmd to null in order
1587                  * to split it, the pmd must remain marked huge at all
1588                  * times or the VM won't take the pmd_trans_huge paths
1589                  * and it won't wait on the anon_vma->root->rwsem to
1590                  * serialize against split_huge_page*.
1591                  */
1592                 pmdp_splitting_flush(vma, address, pmd);
1593                 ret = 1;
1594         }
1595         spin_unlock(&mm->page_table_lock);
1596         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1597
1598         return ret;
1599 }
1600
1601 static void __split_huge_page_refcount(struct page *page,
1602                                        struct list_head *list)
1603 {
1604         int i;
1605         struct zone *zone = page_zone(page);
1606         struct lruvec *lruvec;
1607         int tail_count = 0;
1608
1609         /* prevent PageLRU to go away from under us, and freeze lru stats */
1610         spin_lock_irq(&zone->lru_lock);
1611         lruvec = mem_cgroup_page_lruvec(page, zone);
1612
1613         compound_lock(page);
1614         /* complete memcg works before add pages to LRU */
1615         mem_cgroup_split_huge_fixup(page);
1616
1617         for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
1618                 struct page *page_tail = page + i;
1619
1620                 /* tail_page->_mapcount cannot change */
1621                 BUG_ON(page_mapcount(page_tail) < 0);
1622                 tail_count += page_mapcount(page_tail);
1623                 /* check for overflow */
1624                 BUG_ON(tail_count < 0);
1625                 BUG_ON(atomic_read(&page_tail->_count) != 0);
1626                 /*
1627                  * tail_page->_count is zero and not changing from
1628                  * under us. But get_page_unless_zero() may be running
1629                  * from under us on the tail_page. If we used
1630                  * atomic_set() below instead of atomic_add(), we
1631                  * would then run atomic_set() concurrently with
1632                  * get_page_unless_zero(), and atomic_set() is
1633                  * implemented in C not using locked ops. spin_unlock
1634                  * on x86 sometime uses locked ops because of PPro
1635                  * errata 66, 92, so unless somebody can guarantee
1636                  * atomic_set() here would be safe on all archs (and
1637                  * not only on x86), it's safer to use atomic_add().
1638                  */
1639                 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1640                            &page_tail->_count);
1641
1642                 /* after clearing PageTail the gup refcount can be released */
1643                 smp_mb();
1644
1645                 /*
1646                  * retain hwpoison flag of the poisoned tail page:
1647                  *   fix for the unsuitable process killed on Guest Machine(KVM)
1648                  *   by the memory-failure.
1649                  */
1650                 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON;
1651                 page_tail->flags |= (page->flags &
1652                                      ((1L << PG_referenced) |
1653                                       (1L << PG_swapbacked) |
1654                                       (1L << PG_mlocked) |
1655                                       (1L << PG_uptodate)));
1656                 page_tail->flags |= (1L << PG_dirty);
1657
1658                 /* clear PageTail before overwriting first_page */
1659                 smp_wmb();
1660
1661                 /*
1662                  * __split_huge_page_splitting() already set the
1663                  * splitting bit in all pmd that could map this
1664                  * hugepage, that will ensure no CPU can alter the
1665                  * mapcount on the head page. The mapcount is only
1666                  * accounted in the head page and it has to be
1667                  * transferred to all tail pages in the below code. So
1668                  * for this code to be safe, the split the mapcount
1669                  * can't change. But that doesn't mean userland can't
1670                  * keep changing and reading the page contents while
1671                  * we transfer the mapcount, so the pmd splitting
1672                  * status is achieved setting a reserved bit in the
1673                  * pmd, not by clearing the present bit.
1674                 */
1675                 page_tail->_mapcount = page->_mapcount;
1676
1677                 BUG_ON(page_tail->mapping);
1678                 page_tail->mapping = page->mapping;
1679
1680                 page_tail->index = page->index + i;
1681                 page_nid_xchg_last(page_tail, page_nid_last(page));
1682
1683                 BUG_ON(!PageAnon(page_tail));
1684                 BUG_ON(!PageUptodate(page_tail));
1685                 BUG_ON(!PageDirty(page_tail));
1686                 BUG_ON(!PageSwapBacked(page_tail));
1687
1688                 lru_add_page_tail(page, page_tail, lruvec, list);
1689         }
1690         atomic_sub(tail_count, &page->_count);
1691         BUG_ON(atomic_read(&page->_count) <= 0);
1692
1693         __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1);
1694         __mod_zone_page_state(zone, NR_ANON_PAGES, HPAGE_PMD_NR);
1695
1696         ClearPageCompound(page);
1697         compound_unlock(page);
1698         spin_unlock_irq(&zone->lru_lock);
1699
1700         for (i = 1; i < HPAGE_PMD_NR; i++) {
1701                 struct page *page_tail = page + i;
1702                 BUG_ON(page_count(page_tail) <= 0);
1703                 /*
1704                  * Tail pages may be freed if there wasn't any mapping
1705                  * like if add_to_swap() is running on a lru page that
1706                  * had its mapping zapped. And freeing these pages
1707                  * requires taking the lru_lock so we do the put_page
1708                  * of the tail pages after the split is complete.
1709                  */
1710                 put_page(page_tail);
1711         }
1712
1713         /*
1714          * Only the head page (now become a regular page) is required
1715          * to be pinned by the caller.
1716          */
1717         BUG_ON(page_count(page) <= 0);
1718 }
1719
1720 static int __split_huge_page_map(struct page *page,
1721                                  struct vm_area_struct *vma,
1722                                  unsigned long address)
1723 {
1724         struct mm_struct *mm = vma->vm_mm;
1725         pmd_t *pmd, _pmd;
1726         int ret = 0, i;
1727         pgtable_t pgtable;
1728         unsigned long haddr;
1729
1730         spin_lock(&mm->page_table_lock);
1731         pmd = page_check_address_pmd(page, mm, address,
1732                                      PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG);
1733         if (pmd) {
1734                 pgtable = pgtable_trans_huge_withdraw(mm);
1735                 pmd_populate(mm, &_pmd, pgtable);
1736                 if (pmd_write(*pmd))
1737                         BUG_ON(page_mapcount(page) != 1);
1738
1739                 haddr = address;
1740                 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1741                         pte_t *pte, entry;
1742                         BUG_ON(PageCompound(page+i));
1743                         /*
1744                          * Note that pmd_numa is not transferred deliberately
1745                          * to avoid any possibility that pte_numa leaks to
1746                          * a PROT_NONE VMA by accident.
1747                          */
1748                         entry = mk_pte(page + i, vma->vm_page_prot);
1749                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1750                         if (!pmd_write(*pmd))
1751                                 entry = pte_wrprotect(entry);
1752                         if (!pmd_young(*pmd))
1753                                 entry = pte_mkold(entry);
1754                         pte = pte_offset_map(&_pmd, haddr);
1755                         BUG_ON(!pte_none(*pte));
1756                         set_pte_at(mm, haddr, pte, entry);
1757                         pte_unmap(pte);
1758                 }
1759
1760                 smp_wmb(); /* make pte visible before pmd */
1761                 /*
1762                  * Up to this point the pmd is present and huge and
1763                  * userland has the whole access to the hugepage
1764                  * during the split (which happens in place). If we
1765                  * overwrite the pmd with the not-huge version
1766                  * pointing to the pte here (which of course we could
1767                  * if all CPUs were bug free), userland could trigger
1768                  * a small page size TLB miss on the small sized TLB
1769                  * while the hugepage TLB entry is still established
1770                  * in the huge TLB. Some CPU doesn't like that. See
1771                  * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1772                  * Erratum 383 on page 93. Intel should be safe but is
1773                  * also warns that it's only safe if the permission
1774                  * and cache attributes of the two entries loaded in
1775                  * the two TLB is identical (which should be the case
1776                  * here). But it is generally safer to never allow
1777                  * small and huge TLB entries for the same virtual
1778                  * address to be loaded simultaneously. So instead of
1779                  * doing "pmd_populate(); flush_tlb_range();" we first
1780                  * mark the current pmd notpresent (atomically because
1781                  * here the pmd_trans_huge and pmd_trans_splitting
1782                  * must remain set at all times on the pmd until the
1783                  * split is complete for this pmd), then we flush the
1784                  * SMP TLB and finally we write the non-huge version
1785                  * of the pmd entry with pmd_populate.
1786                  */
1787                 pmdp_invalidate(vma, address, pmd);
1788                 pmd_populate(mm, pmd, pgtable);
1789                 ret = 1;
1790         }
1791         spin_unlock(&mm->page_table_lock);
1792
1793         return ret;
1794 }
1795
1796 /* must be called with anon_vma->root->rwsem held */
1797 static void __split_huge_page(struct page *page,
1798                               struct anon_vma *anon_vma,
1799                               struct list_head *list)
1800 {
1801         int mapcount, mapcount2;
1802         pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1803         struct anon_vma_chain *avc;
1804
1805         BUG_ON(!PageHead(page));
1806         BUG_ON(PageTail(page));
1807
1808         mapcount = 0;
1809         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1810                 struct vm_area_struct *vma = avc->vma;
1811                 unsigned long addr = vma_address(page, vma);
1812                 BUG_ON(is_vma_temporary_stack(vma));
1813                 mapcount += __split_huge_page_splitting(page, vma, addr);
1814         }
1815         /*
1816          * It is critical that new vmas are added to the tail of the
1817          * anon_vma list. This guarantes that if copy_huge_pmd() runs
1818          * and establishes a child pmd before
1819          * __split_huge_page_splitting() freezes the parent pmd (so if
1820          * we fail to prevent copy_huge_pmd() from running until the
1821          * whole __split_huge_page() is complete), we will still see
1822          * the newly established pmd of the child later during the
1823          * walk, to be able to set it as pmd_trans_splitting too.
1824          */
1825         if (mapcount != page_mapcount(page))
1826                 printk(KERN_ERR "mapcount %d page_mapcount %d\n",
1827                        mapcount, page_mapcount(page));
1828         BUG_ON(mapcount != page_mapcount(page));
1829
1830         __split_huge_page_refcount(page, list);
1831
1832         mapcount2 = 0;
1833         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1834                 struct vm_area_struct *vma = avc->vma;
1835                 unsigned long addr = vma_address(page, vma);
1836                 BUG_ON(is_vma_temporary_stack(vma));
1837                 mapcount2 += __split_huge_page_map(page, vma, addr);
1838         }
1839         if (mapcount != mapcount2)
1840                 printk(KERN_ERR "mapcount %d mapcount2 %d page_mapcount %d\n",
1841                        mapcount, mapcount2, page_mapcount(page));
1842         BUG_ON(mapcount != mapcount2);
1843 }
1844
1845 /*
1846  * Split a hugepage into normal pages. This doesn't change the position of head
1847  * page. If @list is null, tail pages will be added to LRU list, otherwise, to
1848  * @list. Both head page and tail pages will inherit mapping, flags, and so on
1849  * from the hugepage.
1850  * Return 0 if the hugepage is split successfully otherwise return 1.
1851  */
1852 int split_huge_page_to_list(struct page *page, struct list_head *list)
1853 {
1854         struct anon_vma *anon_vma;
1855         int ret = 1;
1856
1857         BUG_ON(is_huge_zero_page(page));
1858         BUG_ON(!PageAnon(page));
1859
1860         /*
1861          * The caller does not necessarily hold an mmap_sem that would prevent
1862          * the anon_vma disappearing so we first we take a reference to it
1863          * and then lock the anon_vma for write. This is similar to
1864          * page_lock_anon_vma_read except the write lock is taken to serialise
1865          * against parallel split or collapse operations.
1866          */
1867         anon_vma = page_get_anon_vma(page);
1868         if (!anon_vma)
1869                 goto out;
1870         anon_vma_lock_write(anon_vma);
1871
1872         ret = 0;
1873         if (!PageCompound(page))
1874                 goto out_unlock;
1875
1876         BUG_ON(!PageSwapBacked(page));
1877         __split_huge_page(page, anon_vma, list);
1878         count_vm_event(THP_SPLIT);
1879
1880         BUG_ON(PageCompound(page));
1881 out_unlock:
1882         anon_vma_unlock_write(anon_vma);
1883         put_anon_vma(anon_vma);
1884 out:
1885         return ret;
1886 }
1887
1888 #define VM_NO_THP (VM_SPECIAL|VM_MIXEDMAP|VM_HUGETLB|VM_SHARED|VM_MAYSHARE)
1889
1890 int hugepage_madvise(struct vm_area_struct *vma,
1891                      unsigned long *vm_flags, int advice)
1892 {
1893         struct mm_struct *mm = vma->vm_mm;
1894
1895         switch (advice) {
1896         case MADV_HUGEPAGE:
1897                 /*
1898                  * Be somewhat over-protective like KSM for now!
1899                  */
1900                 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
1901                         return -EINVAL;
1902                 if (mm->def_flags & VM_NOHUGEPAGE)
1903                         return -EINVAL;
1904                 *vm_flags &= ~VM_NOHUGEPAGE;
1905                 *vm_flags |= VM_HUGEPAGE;
1906                 /*
1907                  * If the vma become good for khugepaged to scan,
1908                  * register it here without waiting a page fault that
1909                  * may not happen any time soon.
1910                  */
1911                 if (unlikely(khugepaged_enter_vma_merge(vma)))
1912                         return -ENOMEM;
1913                 break;
1914         case MADV_NOHUGEPAGE:
1915                 /*
1916                  * Be somewhat over-protective like KSM for now!
1917                  */
1918                 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
1919                         return -EINVAL;
1920                 *vm_flags &= ~VM_HUGEPAGE;
1921                 *vm_flags |= VM_NOHUGEPAGE;
1922                 /*
1923                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1924                  * this vma even if we leave the mm registered in khugepaged if
1925                  * it got registered before VM_NOHUGEPAGE was set.
1926                  */
1927                 break;
1928         }
1929
1930         return 0;
1931 }
1932
1933 static int __init khugepaged_slab_init(void)
1934 {
1935         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1936                                           sizeof(struct mm_slot),
1937                                           __alignof__(struct mm_slot), 0, NULL);
1938         if (!mm_slot_cache)
1939                 return -ENOMEM;
1940
1941         return 0;
1942 }
1943
1944 static inline struct mm_slot *alloc_mm_slot(void)
1945 {
1946         if (!mm_slot_cache)     /* initialization failed */
1947                 return NULL;
1948         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1949 }
1950
1951 static inline void free_mm_slot(struct mm_slot *mm_slot)
1952 {
1953         kmem_cache_free(mm_slot_cache, mm_slot);
1954 }
1955
1956 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1957 {
1958         struct mm_slot *mm_slot;
1959
1960         hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
1961                 if (mm == mm_slot->mm)
1962                         return mm_slot;
1963
1964         return NULL;
1965 }
1966
1967 static void insert_to_mm_slots_hash(struct mm_struct *mm,
1968                                     struct mm_slot *mm_slot)
1969 {
1970         mm_slot->mm = mm;
1971         hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
1972 }
1973
1974 static inline int khugepaged_test_exit(struct mm_struct *mm)
1975 {
1976         return atomic_read(&mm->mm_users) == 0;
1977 }
1978
1979 int __khugepaged_enter(struct mm_struct *mm)
1980 {
1981         struct mm_slot *mm_slot;
1982         int wakeup;
1983
1984         mm_slot = alloc_mm_slot();
1985         if (!mm_slot)
1986                 return -ENOMEM;
1987
1988         /* __khugepaged_exit() must not run from under us */
1989         VM_BUG_ON(khugepaged_test_exit(mm));
1990         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
1991                 free_mm_slot(mm_slot);
1992                 return 0;
1993         }
1994
1995         spin_lock(&khugepaged_mm_lock);
1996         insert_to_mm_slots_hash(mm, mm_slot);
1997         /*
1998          * Insert just behind the scanning cursor, to let the area settle
1999          * down a little.
2000          */
2001         wakeup = list_empty(&khugepaged_scan.mm_head);
2002         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
2003         spin_unlock(&khugepaged_mm_lock);
2004
2005         atomic_inc(&mm->mm_count);
2006         if (wakeup)
2007                 wake_up_interruptible(&khugepaged_wait);
2008
2009         return 0;
2010 }
2011
2012 int khugepaged_enter_vma_merge(struct vm_area_struct *vma)
2013 {
2014         unsigned long hstart, hend;
2015         if (!vma->anon_vma)
2016                 /*
2017                  * Not yet faulted in so we will register later in the
2018                  * page fault if needed.
2019                  */
2020                 return 0;
2021         if (vma->vm_ops)
2022                 /* khugepaged not yet working on file or special mappings */
2023                 return 0;
2024         VM_BUG_ON(vma->vm_flags & VM_NO_THP);
2025         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2026         hend = vma->vm_end & HPAGE_PMD_MASK;
2027         if (hstart < hend)
2028                 return khugepaged_enter(vma);
2029         return 0;
2030 }
2031
2032 void __khugepaged_exit(struct mm_struct *mm)
2033 {
2034         struct mm_slot *mm_slot;
2035         int free = 0;
2036
2037         spin_lock(&khugepaged_mm_lock);
2038         mm_slot = get_mm_slot(mm);
2039         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
2040                 hash_del(&mm_slot->hash);
2041                 list_del(&mm_slot->mm_node);
2042                 free = 1;
2043         }
2044         spin_unlock(&khugepaged_mm_lock);
2045
2046         if (free) {
2047                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2048                 free_mm_slot(mm_slot);
2049                 mmdrop(mm);
2050         } else if (mm_slot) {
2051                 /*
2052                  * This is required to serialize against
2053                  * khugepaged_test_exit() (which is guaranteed to run
2054                  * under mmap sem read mode). Stop here (after we
2055                  * return all pagetables will be destroyed) until
2056                  * khugepaged has finished working on the pagetables
2057                  * under the mmap_sem.
2058                  */
2059                 down_write(&mm->mmap_sem);
2060                 up_write(&mm->mmap_sem);
2061         }
2062 }
2063
2064 static void release_pte_page(struct page *page)
2065 {
2066         /* 0 stands for page_is_file_cache(page) == false */
2067         dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
2068         unlock_page(page);
2069         putback_lru_page(page);
2070 }
2071
2072 static void release_pte_pages(pte_t *pte, pte_t *_pte)
2073 {
2074         while (--_pte >= pte) {
2075                 pte_t pteval = *_pte;
2076                 if (!pte_none(pteval))
2077                         release_pte_page(pte_page(pteval));
2078         }
2079 }
2080
2081 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
2082                                         unsigned long address,
2083                                         pte_t *pte)
2084 {
2085         struct page *page;
2086         pte_t *_pte;
2087         int referenced = 0, none = 0;
2088         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
2089              _pte++, address += PAGE_SIZE) {
2090                 pte_t pteval = *_pte;
2091                 if (pte_none(pteval)) {
2092                         if (++none <= khugepaged_max_ptes_none)
2093                                 continue;
2094                         else
2095                                 goto out;
2096                 }
2097                 if (!pte_present(pteval) || !pte_write(pteval))
2098                         goto out;
2099                 page = vm_normal_page(vma, address, pteval);
2100                 if (unlikely(!page))
2101                         goto out;
2102
2103                 VM_BUG_ON(PageCompound(page));
2104                 BUG_ON(!PageAnon(page));
2105                 VM_BUG_ON(!PageSwapBacked(page));
2106
2107                 /* cannot use mapcount: can't collapse if there's a gup pin */
2108                 if (page_count(page) != 1)
2109                         goto out;
2110                 /*
2111                  * We can do it before isolate_lru_page because the
2112                  * page can't be freed from under us. NOTE: PG_lock
2113                  * is needed to serialize against split_huge_page
2114                  * when invoked from the VM.
2115                  */
2116                 if (!trylock_page(page))
2117                         goto out;
2118                 /*
2119                  * Isolate the page to avoid collapsing an hugepage
2120                  * currently in use by the VM.
2121                  */
2122                 if (isolate_lru_page(page)) {
2123                         unlock_page(page);
2124                         goto out;
2125                 }
2126                 /* 0 stands for page_is_file_cache(page) == false */
2127                 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
2128                 VM_BUG_ON(!PageLocked(page));
2129                 VM_BUG_ON(PageLRU(page));
2130
2131                 /* If there is no mapped pte young don't collapse the page */
2132                 if (pte_young(pteval) || PageReferenced(page) ||
2133                     mmu_notifier_test_young(vma->vm_mm, address))
2134                         referenced = 1;
2135         }
2136         if (likely(referenced))
2137                 return 1;
2138 out:
2139         release_pte_pages(pte, _pte);
2140         return 0;
2141 }
2142
2143 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
2144                                       struct vm_area_struct *vma,
2145                                       unsigned long address,
2146                                       spinlock_t *ptl)
2147 {
2148         pte_t *_pte;
2149         for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
2150                 pte_t pteval = *_pte;
2151                 struct page *src_page;
2152
2153                 if (pte_none(pteval)) {
2154                         clear_user_highpage(page, address);
2155                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
2156                 } else {
2157                         src_page = pte_page(pteval);
2158                         copy_user_highpage(page, src_page, address, vma);
2159                         VM_BUG_ON(page_mapcount(src_page) != 1);
2160                         release_pte_page(src_page);
2161                         /*
2162                          * ptl mostly unnecessary, but preempt has to
2163                          * be disabled to update the per-cpu stats
2164                          * inside page_remove_rmap().
2165                          */
2166                         spin_lock(ptl);
2167                         /*
2168                          * paravirt calls inside pte_clear here are
2169                          * superfluous.
2170                          */
2171                         pte_clear(vma->vm_mm, address, _pte);
2172                         page_remove_rmap(src_page);
2173                         spin_unlock(ptl);
2174                         free_page_and_swap_cache(src_page);
2175                 }
2176
2177                 address += PAGE_SIZE;
2178                 page++;
2179         }
2180 }
2181
2182 static void khugepaged_alloc_sleep(void)
2183 {
2184         wait_event_freezable_timeout(khugepaged_wait, false,
2185                         msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2186 }
2187
2188 #ifdef CONFIG_NUMA
2189 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2190 {
2191         if (IS_ERR(*hpage)) {
2192                 if (!*wait)
2193                         return false;
2194
2195                 *wait = false;
2196                 *hpage = NULL;
2197                 khugepaged_alloc_sleep();
2198         } else if (*hpage) {
2199                 put_page(*hpage);
2200                 *hpage = NULL;
2201         }
2202
2203         return true;
2204 }
2205
2206 static struct page
2207 *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
2208                        struct vm_area_struct *vma, unsigned long address,
2209                        int node)
2210 {
2211         VM_BUG_ON(*hpage);
2212         /*
2213          * Allocate the page while the vma is still valid and under
2214          * the mmap_sem read mode so there is no memory allocation
2215          * later when we take the mmap_sem in write mode. This is more
2216          * friendly behavior (OTOH it may actually hide bugs) to
2217          * filesystems in userland with daemons allocating memory in
2218          * the userland I/O paths.  Allocating memory with the
2219          * mmap_sem in read mode is good idea also to allow greater
2220          * scalability.
2221          */
2222         *hpage  = alloc_hugepage_vma(khugepaged_defrag(), vma, address,
2223                                       node, __GFP_OTHER_NODE);
2224
2225         /*
2226          * After allocating the hugepage, release the mmap_sem read lock in
2227          * preparation for taking it in write mode.
2228          */
2229         up_read(&mm->mmap_sem);
2230         if (unlikely(!*hpage)) {
2231                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2232                 *hpage = ERR_PTR(-ENOMEM);
2233                 return NULL;
2234         }
2235
2236         count_vm_event(THP_COLLAPSE_ALLOC);
2237         return *hpage;
2238 }
2239 #else
2240 static struct page *khugepaged_alloc_hugepage(bool *wait)
2241 {
2242         struct page *hpage;
2243
2244         do {
2245                 hpage = alloc_hugepage(khugepaged_defrag());
2246                 if (!hpage) {
2247                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2248                         if (!*wait)
2249                                 return NULL;
2250
2251                         *wait = false;
2252                         khugepaged_alloc_sleep();
2253                 } else
2254                         count_vm_event(THP_COLLAPSE_ALLOC);
2255         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2256
2257         return hpage;
2258 }
2259
2260 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2261 {
2262         if (!*hpage)
2263                 *hpage = khugepaged_alloc_hugepage(wait);
2264
2265         if (unlikely(!*hpage))
2266                 return false;
2267
2268         return true;
2269 }
2270
2271 static struct page
2272 *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
2273                        struct vm_area_struct *vma, unsigned long address,
2274                        int node)
2275 {
2276         up_read(&mm->mmap_sem);
2277         VM_BUG_ON(!*hpage);
2278         return  *hpage;
2279 }
2280 #endif
2281
2282 static bool hugepage_vma_check(struct vm_area_struct *vma)
2283 {
2284         if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2285             (vma->vm_flags & VM_NOHUGEPAGE))
2286                 return false;
2287
2288         if (!vma->anon_vma || vma->vm_ops)
2289                 return false;
2290         if (is_vma_temporary_stack(vma))
2291                 return false;
2292         VM_BUG_ON(vma->vm_flags & VM_NO_THP);
2293         return true;
2294 }
2295
2296 static void collapse_huge_page(struct mm_struct *mm,
2297                                    unsigned long address,
2298                                    struct page **hpage,
2299                                    struct vm_area_struct *vma,
2300                                    int node)
2301 {
2302         pmd_t *pmd, _pmd;
2303         pte_t *pte;
2304         pgtable_t pgtable;
2305         struct page *new_page;
2306         spinlock_t *ptl;
2307         int isolated;
2308         unsigned long hstart, hend;
2309         unsigned long mmun_start;       /* For mmu_notifiers */
2310         unsigned long mmun_end;         /* For mmu_notifiers */
2311
2312         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2313
2314         /* release the mmap_sem read lock. */
2315         new_page = khugepaged_alloc_page(hpage, mm, vma, address, node);
2316         if (!new_page)
2317                 return;
2318
2319         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL)))
2320                 return;
2321
2322         /*
2323          * Prevent all access to pagetables with the exception of
2324          * gup_fast later hanlded by the ptep_clear_flush and the VM
2325          * handled by the anon_vma lock + PG_lock.
2326          */
2327         down_write(&mm->mmap_sem);
2328         if (unlikely(khugepaged_test_exit(mm)))
2329                 goto out;
2330
2331         vma = find_vma(mm, address);
2332         if (!vma)
2333                 goto out;
2334         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2335         hend = vma->vm_end & HPAGE_PMD_MASK;
2336         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
2337                 goto out;
2338         if (!hugepage_vma_check(vma))
2339                 goto out;
2340         pmd = mm_find_pmd(mm, address);
2341         if (!pmd)
2342                 goto out;
2343         if (pmd_trans_huge(*pmd))
2344                 goto out;
2345
2346         anon_vma_lock_write(vma->anon_vma);
2347
2348         pte = pte_offset_map(pmd, address);
2349         ptl = pte_lockptr(mm, pmd);
2350
2351         mmun_start = address;
2352         mmun_end   = address + HPAGE_PMD_SIZE;
2353         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2354         spin_lock(&mm->page_table_lock); /* probably unnecessary */
2355         /*
2356          * After this gup_fast can't run anymore. This also removes
2357          * any huge TLB entry from the CPU so we won't allow
2358          * huge and small TLB entries for the same virtual address
2359          * to avoid the risk of CPU bugs in that area.
2360          */
2361         _pmd = pmdp_clear_flush(vma, address, pmd);
2362         spin_unlock(&mm->page_table_lock);
2363         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2364
2365         spin_lock(ptl);
2366         isolated = __collapse_huge_page_isolate(vma, address, pte);
2367         spin_unlock(ptl);
2368
2369         if (unlikely(!isolated)) {
2370                 pte_unmap(pte);
2371                 spin_lock(&mm->page_table_lock);
2372                 BUG_ON(!pmd_none(*pmd));
2373                 /*
2374                  * We can only use set_pmd_at when establishing
2375                  * hugepmds and never for establishing regular pmds that
2376                  * points to regular pagetables. Use pmd_populate for that
2377                  */
2378                 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
2379                 spin_unlock(&mm->page_table_lock);
2380                 anon_vma_unlock_write(vma->anon_vma);
2381                 goto out;
2382         }
2383
2384         /*
2385          * All pages are isolated and locked so anon_vma rmap
2386          * can't run anymore.
2387          */
2388         anon_vma_unlock_write(vma->anon_vma);
2389
2390         __collapse_huge_page_copy(pte, new_page, vma, address, ptl);
2391         pte_unmap(pte);
2392         __SetPageUptodate(new_page);
2393         pgtable = pmd_pgtable(_pmd);
2394
2395         _pmd = mk_huge_pmd(new_page, vma);
2396
2397         /*
2398          * spin_lock() below is not the equivalent of smp_wmb(), so
2399          * this is needed to avoid the copy_huge_page writes to become
2400          * visible after the set_pmd_at() write.
2401          */
2402         smp_wmb();
2403
2404         spin_lock(&mm->page_table_lock);
2405         BUG_ON(!pmd_none(*pmd));
2406         page_add_new_anon_rmap(new_page, vma, address);
2407         set_pmd_at(mm, address, pmd, _pmd);
2408         update_mmu_cache_pmd(vma, address, pmd);
2409         pgtable_trans_huge_deposit(mm, pgtable);
2410         spin_unlock(&mm->page_table_lock);
2411
2412         *hpage = NULL;
2413
2414         khugepaged_pages_collapsed++;
2415 out_up_write:
2416         up_write(&mm->mmap_sem);
2417         return;
2418
2419 out:
2420         mem_cgroup_uncharge_page(new_page);
2421         goto out_up_write;
2422 }
2423
2424 static int khugepaged_scan_pmd(struct mm_struct *mm,
2425                                struct vm_area_struct *vma,
2426                                unsigned long address,
2427                                struct page **hpage)
2428 {
2429         pmd_t *pmd;
2430         pte_t *pte, *_pte;
2431         int ret = 0, referenced = 0, none = 0;
2432         struct page *page;
2433         unsigned long _address;
2434         spinlock_t *ptl;
2435         int node = NUMA_NO_NODE;
2436
2437         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2438
2439         pmd = mm_find_pmd(mm, address);
2440         if (!pmd)
2441                 goto out;
2442         if (pmd_trans_huge(*pmd))
2443                 goto out;
2444
2445         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2446         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2447              _pte++, _address += PAGE_SIZE) {
2448                 pte_t pteval = *_pte;
2449                 if (pte_none(pteval)) {
2450                         if (++none <= khugepaged_max_ptes_none)
2451                                 continue;
2452                         else
2453                                 goto out_unmap;
2454                 }
2455                 if (!pte_present(pteval) || !pte_write(pteval))
2456                         goto out_unmap;
2457                 page = vm_normal_page(vma, _address, pteval);
2458                 if (unlikely(!page))
2459                         goto out_unmap;
2460                 /*
2461                  * Chose the node of the first page. This could
2462                  * be more sophisticated and look at more pages,
2463                  * but isn't for now.
2464                  */
2465                 if (node == NUMA_NO_NODE)
2466                         node = page_to_nid(page);
2467                 VM_BUG_ON(PageCompound(page));
2468                 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2469                         goto out_unmap;
2470                 /* cannot use mapcount: can't collapse if there's a gup pin */
2471                 if (page_count(page) != 1)
2472                         goto out_unmap;
2473                 if (pte_young(pteval) || PageReferenced(page) ||
2474                     mmu_notifier_test_young(vma->vm_mm, address))
2475                         referenced = 1;
2476         }
2477         if (referenced)
2478                 ret = 1;
2479 out_unmap:
2480         pte_unmap_unlock(pte, ptl);
2481         if (ret)
2482                 /* collapse_huge_page will return with the mmap_sem released */
2483                 collapse_huge_page(mm, address, hpage, vma, node);
2484 out:
2485         return ret;
2486 }
2487
2488 static void collect_mm_slot(struct mm_slot *mm_slot)
2489 {
2490         struct mm_struct *mm = mm_slot->mm;
2491
2492         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2493
2494         if (khugepaged_test_exit(mm)) {
2495                 /* free mm_slot */
2496                 hash_del(&mm_slot->hash);
2497                 list_del(&mm_slot->mm_node);
2498
2499                 /*
2500                  * Not strictly needed because the mm exited already.
2501                  *
2502                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2503                  */
2504
2505                 /* khugepaged_mm_lock actually not necessary for the below */
2506                 free_mm_slot(mm_slot);
2507                 mmdrop(mm);
2508         }
2509 }
2510
2511 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2512                                             struct page **hpage)
2513         __releases(&khugepaged_mm_lock)
2514         __acquires(&khugepaged_mm_lock)
2515 {
2516         struct mm_slot *mm_slot;
2517         struct mm_struct *mm;
2518         struct vm_area_struct *vma;
2519         int progress = 0;
2520
2521         VM_BUG_ON(!pages);
2522         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2523
2524         if (khugepaged_scan.mm_slot)
2525                 mm_slot = khugepaged_scan.mm_slot;
2526         else {
2527                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2528                                      struct mm_slot, mm_node);
2529                 khugepaged_scan.address = 0;
2530                 khugepaged_scan.mm_slot = mm_slot;
2531         }
2532         spin_unlock(&khugepaged_mm_lock);
2533
2534         mm = mm_slot->mm;
2535         down_read(&mm->mmap_sem);
2536         if (unlikely(khugepaged_test_exit(mm)))
2537                 vma = NULL;
2538         else
2539                 vma = find_vma(mm, khugepaged_scan.address);
2540
2541         progress++;
2542         for (; vma; vma = vma->vm_next) {
2543                 unsigned long hstart, hend;
2544
2545                 cond_resched();
2546                 if (unlikely(khugepaged_test_exit(mm))) {
2547                         progress++;
2548                         break;
2549                 }
2550                 if (!hugepage_vma_check(vma)) {
2551 skip:
2552                         progress++;
2553                         continue;
2554                 }
2555                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2556                 hend = vma->vm_end & HPAGE_PMD_MASK;
2557                 if (hstart >= hend)
2558                         goto skip;
2559                 if (khugepaged_scan.address > hend)
2560                         goto skip;
2561                 if (khugepaged_scan.address < hstart)
2562                         khugepaged_scan.address = hstart;
2563                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2564
2565                 while (khugepaged_scan.address < hend) {
2566                         int ret;
2567                         cond_resched();
2568                         if (unlikely(khugepaged_test_exit(mm)))
2569                                 goto breakouterloop;
2570
2571                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2572                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2573                                   hend);
2574                         ret = khugepaged_scan_pmd(mm, vma,
2575                                                   khugepaged_scan.address,
2576                                                   hpage);
2577                         /* move to next address */
2578                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2579                         progress += HPAGE_PMD_NR;
2580                         if (ret)
2581                                 /* we released mmap_sem so break loop */
2582                                 goto breakouterloop_mmap_sem;
2583                         if (progress >= pages)
2584                                 goto breakouterloop;
2585                 }
2586         }
2587 breakouterloop:
2588         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2589 breakouterloop_mmap_sem:
2590
2591         spin_lock(&khugepaged_mm_lock);
2592         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2593         /*
2594          * Release the current mm_slot if this mm is about to die, or
2595          * if we scanned all vmas of this mm.
2596          */
2597         if (khugepaged_test_exit(mm) || !vma) {
2598                 /*
2599                  * Make sure that if mm_users is reaching zero while
2600                  * khugepaged runs here, khugepaged_exit will find
2601                  * mm_slot not pointing to the exiting mm.
2602                  */
2603                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2604                         khugepaged_scan.mm_slot = list_entry(
2605                                 mm_slot->mm_node.next,
2606                                 struct mm_slot, mm_node);
2607                         khugepaged_scan.address = 0;
2608                 } else {
2609                         khugepaged_scan.mm_slot = NULL;
2610                         khugepaged_full_scans++;
2611                 }
2612
2613                 collect_mm_slot(mm_slot);
2614         }
2615
2616         return progress;
2617 }
2618
2619 static int khugepaged_has_work(void)
2620 {
2621         return !list_empty(&khugepaged_scan.mm_head) &&
2622                 khugepaged_enabled();
2623 }
2624
2625 static int khugepaged_wait_event(void)
2626 {
2627         return !list_empty(&khugepaged_scan.mm_head) ||
2628                 kthread_should_stop();
2629 }
2630
2631 static void khugepaged_do_scan(void)
2632 {
2633         struct page *hpage = NULL;
2634         unsigned int progress = 0, pass_through_head = 0;
2635         unsigned int pages = khugepaged_pages_to_scan;
2636         bool wait = true;
2637
2638         barrier(); /* write khugepaged_pages_to_scan to local stack */
2639
2640         while (progress < pages) {
2641                 if (!khugepaged_prealloc_page(&hpage, &wait))
2642                         break;
2643
2644                 cond_resched();
2645
2646                 if (unlikely(kthread_should_stop() || freezing(current)))
2647                         break;
2648
2649                 spin_lock(&khugepaged_mm_lock);
2650                 if (!khugepaged_scan.mm_slot)
2651                         pass_through_head++;
2652                 if (khugepaged_has_work() &&
2653                     pass_through_head < 2)
2654                         progress += khugepaged_scan_mm_slot(pages - progress,
2655                                                             &hpage);
2656                 else
2657                         progress = pages;
2658                 spin_unlock(&khugepaged_mm_lock);
2659         }
2660
2661         if (!IS_ERR_OR_NULL(hpage))
2662                 put_page(hpage);
2663 }
2664
2665 static void khugepaged_wait_work(void)
2666 {
2667         try_to_freeze();
2668
2669         if (khugepaged_has_work()) {
2670                 if (!khugepaged_scan_sleep_millisecs)
2671                         return;
2672
2673                 wait_event_freezable_timeout(khugepaged_wait,
2674                                              kthread_should_stop(),
2675                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2676                 return;
2677         }
2678
2679         if (khugepaged_enabled())
2680                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2681 }
2682
2683 static int khugepaged(void *none)
2684 {
2685         struct mm_slot *mm_slot;
2686
2687         set_freezable();
2688         set_user_nice(current, 19);
2689
2690         while (!kthread_should_stop()) {
2691                 khugepaged_do_scan();
2692                 khugepaged_wait_work();
2693         }
2694
2695         spin_lock(&khugepaged_mm_lock);
2696         mm_slot = khugepaged_scan.mm_slot;
2697         khugepaged_scan.mm_slot = NULL;
2698         if (mm_slot)
2699                 collect_mm_slot(mm_slot);
2700         spin_unlock(&khugepaged_mm_lock);
2701         return 0;
2702 }
2703
2704 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2705                 unsigned long haddr, pmd_t *pmd)
2706 {
2707         struct mm_struct *mm = vma->vm_mm;
2708         pgtable_t pgtable;
2709         pmd_t _pmd;
2710         int i;
2711
2712         pmdp_clear_flush(vma, haddr, pmd);
2713         /* leave pmd empty until pte is filled */
2714
2715         pgtable = pgtable_trans_huge_withdraw(mm);
2716         pmd_populate(mm, &_pmd, pgtable);
2717
2718         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2719                 pte_t *pte, entry;
2720                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2721                 entry = pte_mkspecial(entry);
2722                 pte = pte_offset_map(&_pmd, haddr);
2723                 VM_BUG_ON(!pte_none(*pte));
2724                 set_pte_at(mm, haddr, pte, entry);
2725                 pte_unmap(pte);
2726         }
2727         smp_wmb(); /* make pte visible before pmd */
2728         pmd_populate(mm, pmd, pgtable);
2729         put_huge_zero_page();
2730 }
2731
2732 void __split_huge_page_pmd(struct vm_area_struct *vma, unsigned long address,
2733                 pmd_t *pmd)
2734 {
2735         struct page *page;
2736         struct mm_struct *mm = vma->vm_mm;
2737         unsigned long haddr = address & HPAGE_PMD_MASK;
2738         unsigned long mmun_start;       /* For mmu_notifiers */
2739         unsigned long mmun_end;         /* For mmu_notifiers */
2740
2741         BUG_ON(vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE);
2742
2743         mmun_start = haddr;
2744         mmun_end   = haddr + HPAGE_PMD_SIZE;
2745 again:
2746         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2747         spin_lock(&mm->page_table_lock);
2748         if (unlikely(!pmd_trans_huge(*pmd))) {
2749                 spin_unlock(&mm->page_table_lock);
2750                 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2751                 return;
2752         }
2753         if (is_huge_zero_pmd(*pmd)) {
2754                 __split_huge_zero_page_pmd(vma, haddr, pmd);
2755                 spin_unlock(&mm->page_table_lock);
2756                 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2757                 return;
2758         }
2759         page = pmd_page(*pmd);
2760         VM_BUG_ON(!page_count(page));
2761         get_page(page);
2762         spin_unlock(&mm->page_table_lock);
2763         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2764
2765         split_huge_page(page);
2766
2767         put_page(page);
2768
2769         /*
2770          * We don't always have down_write of mmap_sem here: a racing
2771          * do_huge_pmd_wp_page() might have copied-on-write to another
2772          * huge page before our split_huge_page() got the anon_vma lock.
2773          */
2774         if (unlikely(pmd_trans_huge(*pmd)))
2775                 goto again;
2776 }
2777
2778 void split_huge_page_pmd_mm(struct mm_struct *mm, unsigned long address,
2779                 pmd_t *pmd)
2780 {
2781         struct vm_area_struct *vma;
2782
2783         vma = find_vma(mm, address);
2784         BUG_ON(vma == NULL);
2785         split_huge_page_pmd(vma, address, pmd);
2786 }
2787
2788 static void split_huge_page_address(struct mm_struct *mm,
2789                                     unsigned long address)
2790 {
2791         pmd_t *pmd;
2792
2793         VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2794
2795         pmd = mm_find_pmd(mm, address);
2796         if (!pmd)
2797                 return;
2798         /*
2799          * Caller holds the mmap_sem write mode, so a huge pmd cannot
2800          * materialize from under us.
2801          */
2802         split_huge_page_pmd_mm(mm, address, pmd);
2803 }
2804
2805 void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2806                              unsigned long start,
2807                              unsigned long end,
2808                              long adjust_next)
2809 {
2810         /*
2811          * If the new start address isn't hpage aligned and it could
2812          * previously contain an hugepage: check if we need to split
2813          * an huge pmd.
2814          */
2815         if (start & ~HPAGE_PMD_MASK &&
2816             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2817             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2818                 split_huge_page_address(vma->vm_mm, start);
2819
2820         /*
2821          * If the new end address isn't hpage aligned and it could
2822          * previously contain an hugepage: check if we need to split
2823          * an huge pmd.
2824          */
2825         if (end & ~HPAGE_PMD_MASK &&
2826             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2827             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2828                 split_huge_page_address(vma->vm_mm, end);
2829
2830         /*
2831          * If we're also updating the vma->vm_next->vm_start, if the new
2832          * vm_next->vm_start isn't page aligned and it could previously
2833          * contain an hugepage: check if we need to split an huge pmd.
2834          */
2835         if (adjust_next > 0) {
2836                 struct vm_area_struct *next = vma->vm_next;
2837                 unsigned long nstart = next->vm_start;
2838                 nstart += adjust_next << PAGE_SHIFT;
2839                 if (nstart & ~HPAGE_PMD_MASK &&
2840                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2841                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2842                         split_huge_page_address(next->vm_mm, nstart);
2843         }
2844 }